Skip to content
This repository was archived by the owner on Mar 28, 2020. It is now read-only.

Commit fd5fafb

Browse files
topperctopperc
committed
[X86] Fix copy paste mistake in vector-idiv-v2i32.ll. Add missing test case.
Some of the test cases contained the same load twice instead of a different load. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@340833 91177308-0d34-0410-b5e6-96231b3b80d8
1 parent da360ea commit fd5fafb

File tree

1 file changed

+169
-61
lines changed

1 file changed

+169
-61
lines changed

‎test/CodeGen/X86/vector-idiv-v2i32.ll‎

Lines changed: 169 additions & 61 deletions
Original file line numberDiff line numberDiff line change
@@ -437,21 +437,99 @@ define void @test_srem_pow2_v2i32(<2 x i32>* %x, <2 x i32>* %y) nounwind {
437437
ret void
438438
}
439439

440+
define void @test_udiv_v2i32(<2 x i32>* %x, <2 x i32>* %y, <2 x i32>* %z) nounwind {
441+
; X64-LABEL: test_udiv_v2i32:
442+
; X64: # %bb.0:
443+
; X64-NEXT: movq %rdx, %rcx
444+
; X64-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
445+
; X64-NEXT: pxor %xmm1, %xmm1
446+
; X64-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
447+
; X64-NEXT: movq {{.*#+}} xmm2 = mem[0],zero
448+
; X64-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
449+
; X64-NEXT: movq %xmm0, %rax
450+
; X64-NEXT: movq %xmm2, %rsi
451+
; X64-NEXT: xorl %edx, %edx
452+
; X64-NEXT: divq %rsi
453+
; X64-NEXT: movq %rax, %xmm1
454+
; X64-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
455+
; X64-NEXT: movq %xmm0, %rax
456+
; X64-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,0,1]
457+
; X64-NEXT: movq %xmm0, %rsi
458+
; X64-NEXT: xorl %edx, %edx
459+
; X64-NEXT: divq %rsi
460+
; X64-NEXT: movq %rax, %xmm0
461+
; X64-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
462+
; X64-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
463+
; X64-NEXT: movq %xmm0, (%rcx)
464+
; X64-NEXT: retq
465+
;
466+
; X86-LABEL: test_udiv_v2i32:
467+
; X86: # %bb.0:
468+
; X86-NEXT: pushl %esi
469+
; X86-NEXT: subl $56, %esp
470+
; X86-NEXT: movl {{[0-9]+}}(%esp), %esi
471+
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
472+
; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
473+
; X86-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
474+
; X86-NEXT: pxor %xmm1, %xmm1
475+
; X86-NEXT: movdqa %xmm0, %xmm2
476+
; X86-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
477+
; X86-NEXT: movdqu %xmm2, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
478+
; X86-NEXT: movq {{.*#+}} xmm2 = mem[0],zero
479+
; X86-NEXT: movdqa %xmm2, %xmm3
480+
; X86-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]
481+
; X86-NEXT: movd %xmm2, {{[0-9]+}}(%esp)
482+
; X86-NEXT: movd %xmm0, (%esp)
483+
; X86-NEXT: movl $0, {{[0-9]+}}(%esp)
484+
; X86-NEXT: movl $0, {{[0-9]+}}(%esp)
485+
; X86-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,0,1]
486+
; X86-NEXT: movdqu %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
487+
; X86-NEXT: calll __udivdi3
488+
; X86-NEXT: movups {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 # 16-byte Reload
489+
; X86-NEXT: movss %xmm0, {{[0-9]+}}(%esp)
490+
; X86-NEXT: movdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 # 16-byte Reload
491+
; X86-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
492+
; X86-NEXT: movd %xmm0, (%esp)
493+
; X86-NEXT: movl $0, {{[0-9]+}}(%esp)
494+
; X86-NEXT: movl $0, {{[0-9]+}}(%esp)
495+
; X86-NEXT: movd %eax, %xmm0
496+
; X86-NEXT: movdqu %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
497+
; X86-NEXT: calll __udivdi3
498+
; X86-NEXT: movd %eax, %xmm0
499+
; X86-NEXT: movdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 # 16-byte Reload
500+
; X86-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
501+
; X86-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
502+
; X86-NEXT: movq %xmm0, (%esi)
503+
; X86-NEXT: addl $56, %esp
504+
; X86-NEXT: popl %esi
505+
; X86-NEXT: retl
506+
%a = load <2 x i32>, <2 x i32>* %x
507+
%b = load <2 x i32>, <2 x i32>* %y
508+
%c = udiv <2 x i32> %a, %b
509+
store <2 x i32> %c, <2 x i32>* %z
510+
ret void
511+
}
512+
440513
define void @test_urem_v2i32(<2 x i32>* %x, <2 x i32>* %y, <2 x i32>* %z) nounwind {
441514
; X64-LABEL: test_urem_v2i32:
442515
; X64: # %bb.0:
443516
; X64-NEXT: movq %rdx, %rcx
444517
; X64-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
445518
; X64-NEXT: pxor %xmm1, %xmm1
446519
; X64-NEXT: punpckldq {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
520+
; X64-NEXT: movq {{.*#+}} xmm2 = mem[0],zero
521+
; X64-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
447522
; X64-NEXT: movq %xmm0, %rax
523+
; X64-NEXT: movq %xmm2, %rsi
448524
; X64-NEXT: xorl %edx, %edx
449-
; X64-NEXT: divq %rax
525+
; X64-NEXT: divq %rsi
450526
; X64-NEXT: movq %rdx, %xmm1
451527
; X64-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
452528
; X64-NEXT: movq %xmm0, %rax
529+
; X64-NEXT: pshufd {{.*#+}} xmm0 = xmm2[2,3,0,1]
530+
; X64-NEXT: movq %xmm0, %rsi
453531
; X64-NEXT: xorl %edx, %edx
454-
; X64-NEXT: divq %rax
532+
; X64-NEXT: divq %rsi
455533
; X64-NEXT: movq %rdx, %xmm0
456534
; X64-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
457535
; X64-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
@@ -461,20 +539,30 @@ define void @test_urem_v2i32(<2 x i32>* %x, <2 x i32>* %y, <2 x i32>* %z) nounwi
461539
; X86-LABEL: test_urem_v2i32:
462540
; X86: # %bb.0:
463541
; X86-NEXT: pushl %esi
464-
; X86-NEXT: subl $40, %esp
542+
; X86-NEXT: subl $56, %esp
465543
; X86-NEXT: movl {{[0-9]+}}(%esp), %esi
466544
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
467-
; X86-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
468-
; X86-NEXT: movups %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
469-
; X86-NEXT: movss %xmm0, {{[0-9]+}}(%esp)
470-
; X86-NEXT: movss %xmm0, (%esp)
545+
; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
546+
; X86-NEXT: movq {{.*#+}} xmm0 = mem[0],zero
547+
; X86-NEXT: pxor %xmm1, %xmm1
548+
; X86-NEXT: movdqa %xmm0, %xmm2
549+
; X86-NEXT: punpckldq {{.*#+}} xmm2 = xmm2[0],xmm1[0],xmm2[1],xmm1[1]
550+
; X86-NEXT: movdqu %xmm2, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
551+
; X86-NEXT: movq {{.*#+}} xmm2 = mem[0],zero
552+
; X86-NEXT: movdqa %xmm2, %xmm3
553+
; X86-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]
554+
; X86-NEXT: movd %xmm2, {{[0-9]+}}(%esp)
555+
; X86-NEXT: movd %xmm0, (%esp)
471556
; X86-NEXT: movl $0, {{[0-9]+}}(%esp)
472557
; X86-NEXT: movl $0, {{[0-9]+}}(%esp)
558+
; X86-NEXT: pshufd {{.*#+}} xmm0 = xmm3[2,3,0,1]
559+
; X86-NEXT: movdqu %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
473560
; X86-NEXT: calll __umoddi3
474561
; X86-NEXT: movups {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 # 16-byte Reload
475-
; X86-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,0,1]
476562
; X86-NEXT: movss %xmm0, {{[0-9]+}}(%esp)
477-
; X86-NEXT: movss %xmm0, (%esp)
563+
; X86-NEXT: movdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm0 # 16-byte Reload
564+
; X86-NEXT: pshufd {{.*#+}} xmm0 = xmm0[2,3,0,1]
565+
; X86-NEXT: movd %xmm0, (%esp)
478566
; X86-NEXT: movl $0, {{[0-9]+}}(%esp)
479567
; X86-NEXT: movl $0, {{[0-9]+}}(%esp)
480568
; X86-NEXT: movd %eax, %xmm0
@@ -485,11 +573,11 @@ define void @test_urem_v2i32(<2 x i32>* %x, <2 x i32>* %y, <2 x i32>* %z) nounwi
485573
; X86-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
486574
; X86-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
487575
; X86-NEXT: movq %xmm0, (%esi)
488-
; X86-NEXT: addl $40, %esp
576+
; X86-NEXT: addl $56, %esp
489577
; X86-NEXT: popl %esi
490578
; X86-NEXT: retl
491579
%a = load <2 x i32>, <2 x i32>* %x
492-
%b = load <2 x i32>, <2 x i32>* %x
580+
%b = load <2 x i32>, <2 x i32>* %y
493581
%c = urem <2 x i32> %a, %b
494582
store <2 x i32> %c, <2 x i32>* %z
495583
ret void
@@ -498,62 +586,72 @@ define void @test_urem_v2i32(<2 x i32>* %x, <2 x i32>* %y, <2 x i32>* %z) nounwi
498586
define void @test_sdiv_v2i32(<2 x i32>* %x, <2 x i32>* %y, <2 x i32>* %z) nounwind {
499587
; X64-LABEL: test_sdiv_v2i32:
500588
; X64: # %bb.0:
501-
; X64-NEXT: movq %rdx, %rcx
502-
; X64-NEXT: movslq (%rdi), %rsi
503-
; X64-NEXT: movslq 4(%rdi), %rdi
504-
; X64-NEXT: movq %rdi, %rax
589+
; X64-NEXT: movq %rdx, %r8
590+
; X64-NEXT: movslq (%rdi), %rcx
591+
; X64-NEXT: movslq 4(%rdi), %rax
592+
; X64-NEXT: movslq (%rsi), %rdi
593+
; X64-NEXT: movslq 4(%rsi), %rsi
505594
; X64-NEXT: cqto
506-
; X64-NEXT: idivq %rdi
595+
; X64-NEXT: idivq %rsi
507596
; X64-NEXT: movq %rax, %xmm0
508-
; X64-NEXT: movq %rsi, %rax
597+
; X64-NEXT: movq %rcx, %rax
509598
; X64-NEXT: cqto
510-
; X64-NEXT: idivq %rsi
599+
; X64-NEXT: idivq %rdi
511600
; X64-NEXT: movq %rax, %xmm1
512601
; X64-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
513602
; X64-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
514-
; X64-NEXT: movq %xmm0, (%rcx)
603+
; X64-NEXT: movq %xmm0, (%r8)
515604
; X64-NEXT: retq
516605
;
517606
; X86-LABEL: test_sdiv_v2i32:
518607
; X86: # %bb.0:
608+
; X86-NEXT: pushl %ebp
519609
; X86-NEXT: pushl %ebx
520610
; X86-NEXT: pushl %edi
521611
; X86-NEXT: pushl %esi
522-
; X86-NEXT: subl $16, %esp
523-
; X86-NEXT: movl {{[0-9]+}}(%esp), %esi
612+
; X86-NEXT: subl $44, %esp
524613
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
525-
; X86-NEXT: movl (%eax), %edi
526-
; X86-NEXT: movl 4(%eax), %eax
527-
; X86-NEXT: movl %edi, %ebx
528-
; X86-NEXT: sarl $31, %ebx
529-
; X86-NEXT: movl %eax, %ecx
530-
; X86-NEXT: sarl $31, %ecx
531-
; X86-NEXT: pushl %ecx
614+
; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
615+
; X86-NEXT: movl (%ecx), %edi
616+
; X86-NEXT: movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
617+
; X86-NEXT: movl 4(%ecx), %esi
618+
; X86-NEXT: sarl $31, %edi
619+
; X86-NEXT: movl %esi, %edx
620+
; X86-NEXT: sarl $31, %edx
621+
; X86-NEXT: movl (%eax), %ebx
622+
; X86-NEXT: movl 4(%eax), %ecx
623+
; X86-NEXT: movl %ebx, %ebp
624+
; X86-NEXT: sarl $31, %ebp
625+
; X86-NEXT: movl %ecx, %eax
626+
; X86-NEXT: sarl $31, %eax
532627
; X86-NEXT: pushl %eax
533628
; X86-NEXT: pushl %ecx
534-
; X86-NEXT: pushl %eax
629+
; X86-NEXT: pushl %edx
630+
; X86-NEXT: pushl %esi
535631
; X86-NEXT: calll __divdi3
536632
; X86-NEXT: addl $16, %esp
537633
; X86-NEXT: movd %eax, %xmm0
538-
; X86-NEXT: movdqu %xmm0, (%esp) # 16-byte Spill
539-
; X86-NEXT: pushl %ebx
540-
; X86-NEXT: pushl %edi
634+
; X86-NEXT: movdqu %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
635+
; X86-NEXT: pushl %ebp
541636
; X86-NEXT: pushl %ebx
542637
; X86-NEXT: pushl %edi
638+
; X86-NEXT: pushl {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Reload
543639
; X86-NEXT: calll __divdi3
544640
; X86-NEXT: addl $16, %esp
545641
; X86-NEXT: movd %eax, %xmm0
546-
; X86-NEXT: movdqu (%esp), %xmm1 # 16-byte Reload
642+
; X86-NEXT: movdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 # 16-byte Reload
547643
; X86-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
548644
; X86-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
549-
; X86-NEXT: movq %xmm0, (%esi)
550-
; X86-NEXT: addl $16, %esp
645+
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
646+
; X86-NEXT: movq %xmm0, (%eax)
647+
; X86-NEXT: addl $44, %esp
551648
; X86-NEXT: popl %esi
552649
; X86-NEXT: popl %edi
553650
; X86-NEXT: popl %ebx
651+
; X86-NEXT: popl %ebp
554652
; X86-NEXT: retl
555653
%a = load <2 x i32>, <2 x i32>* %x
556-
%b = load <2 x i32>, <2 x i32>* %x
654+
%b = load <2 x i32>, <2 x i32>* %y
557655
%c = sdiv <2 x i32> %a, %b
558656
store <2 x i32> %c, <2 x i32>* %z
559657
ret void
@@ -562,62 +660,72 @@ define void @test_sdiv_v2i32(<2 x i32>* %x, <2 x i32>* %y, <2 x i32>* %z) nounwi
562660
define void @test_srem_v2i32(<2 x i32>* %x, <2 x i32>* %y, <2 x i32>* %z) nounwind {
563661
; X64-LABEL: test_srem_v2i32:
564662
; X64: # %bb.0:
565-
; X64-NEXT: movq %rdx, %rcx
566-
; X64-NEXT: movslq (%rdi), %rsi
567-
; X64-NEXT: movslq 4(%rdi), %rdi
568-
; X64-NEXT: movq %rdi, %rax
663+
; X64-NEXT: movq %rdx, %r8
664+
; X64-NEXT: movslq (%rdi), %rcx
665+
; X64-NEXT: movslq 4(%rdi), %rax
666+
; X64-NEXT: movslq (%rsi), %rdi
667+
; X64-NEXT: movslq 4(%rsi), %rsi
569668
; X64-NEXT: cqto
570-
; X64-NEXT: idivq %rdi
669+
; X64-NEXT: idivq %rsi
571670
; X64-NEXT: movq %rax, %xmm0
572-
; X64-NEXT: movq %rsi, %rax
671+
; X64-NEXT: movq %rcx, %rax
573672
; X64-NEXT: cqto
574-
; X64-NEXT: idivq %rsi
673+
; X64-NEXT: idivq %rdi
575674
; X64-NEXT: movq %rax, %xmm1
576675
; X64-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
577676
; X64-NEXT: pshufd {{.*#+}} xmm0 = xmm1[0,2,2,3]
578-
; X64-NEXT: movq %xmm0, (%rcx)
677+
; X64-NEXT: movq %xmm0, (%r8)
579678
; X64-NEXT: retq
580679
;
581680
; X86-LABEL: test_srem_v2i32:
582681
; X86: # %bb.0:
682+
; X86-NEXT: pushl %ebp
583683
; X86-NEXT: pushl %ebx
584684
; X86-NEXT: pushl %edi
585685
; X86-NEXT: pushl %esi
586-
; X86-NEXT: subl $16, %esp
587-
; X86-NEXT: movl {{[0-9]+}}(%esp), %esi
686+
; X86-NEXT: subl $44, %esp
588687
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
589-
; X86-NEXT: movl (%eax), %edi
590-
; X86-NEXT: movl 4(%eax), %eax
591-
; X86-NEXT: movl %edi, %ebx
592-
; X86-NEXT: sarl $31, %ebx
593-
; X86-NEXT: movl %eax, %ecx
594-
; X86-NEXT: sarl $31, %ecx
595-
; X86-NEXT: pushl %ecx
688+
; X86-NEXT: movl {{[0-9]+}}(%esp), %ecx
689+
; X86-NEXT: movl (%ecx), %edi
690+
; X86-NEXT: movl %edi, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
691+
; X86-NEXT: movl 4(%ecx), %esi
692+
; X86-NEXT: sarl $31, %edi
693+
; X86-NEXT: movl %esi, %edx
694+
; X86-NEXT: sarl $31, %edx
695+
; X86-NEXT: movl (%eax), %ebx
696+
; X86-NEXT: movl 4(%eax), %ecx
697+
; X86-NEXT: movl %ebx, %ebp
698+
; X86-NEXT: sarl $31, %ebp
699+
; X86-NEXT: movl %ecx, %eax
700+
; X86-NEXT: sarl $31, %eax
596701
; X86-NEXT: pushl %eax
597702
; X86-NEXT: pushl %ecx
598-
; X86-NEXT: pushl %eax
703+
; X86-NEXT: pushl %edx
704+
; X86-NEXT: pushl %esi
599705
; X86-NEXT: calll __divdi3
600706
; X86-NEXT: addl $16, %esp
601707
; X86-NEXT: movd %eax, %xmm0
602-
; X86-NEXT: movdqu %xmm0, (%esp) # 16-byte Spill
603-
; X86-NEXT: pushl %ebx
604-
; X86-NEXT: pushl %edi
708+
; X86-NEXT: movdqu %xmm0, {{[-0-9]+}}(%e{{[sb]}}p) # 16-byte Spill
709+
; X86-NEXT: pushl %ebp
605710
; X86-NEXT: pushl %ebx
606711
; X86-NEXT: pushl %edi
712+
; X86-NEXT: pushl {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Folded Reload
607713
; X86-NEXT: calll __divdi3
608714
; X86-NEXT: addl $16, %esp
609715
; X86-NEXT: movd %eax, %xmm0
610-
; X86-NEXT: movdqu (%esp), %xmm1 # 16-byte Reload
716+
; X86-NEXT: movdqu {{[-0-9]+}}(%e{{[sb]}}p), %xmm1 # 16-byte Reload
611717
; X86-NEXT: punpcklqdq {{.*#+}} xmm0 = xmm0[0],xmm1[0]
612718
; X86-NEXT: pshufd {{.*#+}} xmm0 = xmm0[0,2,2,3]
613-
; X86-NEXT: movq %xmm0, (%esi)
614-
; X86-NEXT: addl $16, %esp
719+
; X86-NEXT: movl {{[0-9]+}}(%esp), %eax
720+
; X86-NEXT: movq %xmm0, (%eax)
721+
; X86-NEXT: addl $44, %esp
615722
; X86-NEXT: popl %esi
616723
; X86-NEXT: popl %edi
617724
; X86-NEXT: popl %ebx
725+
; X86-NEXT: popl %ebp
618726
; X86-NEXT: retl
619727
%a = load <2 x i32>, <2 x i32>* %x
620-
%b = load <2 x i32>, <2 x i32>* %x
728+
%b = load <2 x i32>, <2 x i32>* %y
621729
%c = sdiv <2 x i32> %a, %b
622730
store <2 x i32> %c, <2 x i32>* %z
623731
ret void

0 commit comments

Comments
 (0)